Interscript
The Interscript dataset contains interactive user feedback on a T5-11B model generated scripts.
Dataset
- data.json contains the data in an easy to read JSON format. data.jsonl contains the data in a JSONL format. The file contains 8466 samples, one sample per line. Every sample is a JSON object with the following fields:
{
"input_script": "push chair in -> pull chair in; pull chair in -> push chair against wall; push chair against wall -> straighten chair legs; straighten chair legs -> Push all chairs in; line up the chairs -> push chair in",
"input_feedback": "One would not pull chair in if they had initially pushed it in.",
"output_script": "push chair against wall -> straighten chair legs;straighten chair legs -> Push all chairs in;line up the chairs -> push chair in;push chair in -> push chair against wall",
"metadata": {
"id": "301KG0KX9BKTC0HB7Z9SV1Y5HAFH2Y.2_implicit.gp",
"goal": "push all chairs in",
"is_distractor": false,
"feedback_type": "implicit.gp",
"edit": "Remove node 'pull chair in'",
"input_script_formatted": [
"1. line up the chairs",
"2. push chair in",
"3. pull chair in",
"4. push chair against wall",
"5. straighten chair legs",
"6. Push all chairs in"
],
"output_script_formatted": [
"1. line up the chairs",
"2. push chair in",
"3. push chair against wall",
"4. straighten chair legs",
"5. Push all chairs in"
]
}
}
The description of the fields is as follows:
input_script
: Model generated script $y_{bad}$.input_feedback
: User feedback on the input script $f$.output_script
: Fixed output script $y_{good}$.
Metadata contains additional information about the sample. Some important fields are:
id
: Unique identifier of the sample.goal
: Goal of the script.is_distractor
: Whether the feedback is a distractor (please see Section 4 for more details).feedback_type
: Type of feedback (please see Section 4 "Annotation" for more details).edit
: Theinput_feedback
presented as an edit operation on the input script, that is, the edit operation that transforms the input script into the output script.input_script_formatted
: The input script presented as a list of sentences.output_script_formatted
: The output script presented as a list of sentences.
Data collection process
- We use Amazon Mechanical Turk to collect feedback on erroneous scripts from users.
- An overview of the process is captured in the following figure:
Amazon Mechanical Turk Template
- turk_template.html contains the template for Amazon Mechanical Turk HITs.